[ROCm][DSV4][Perf] Optimize Triton sparse-MLA decode on gfx950 (#52212)

Signed-off-by: Fangzhou Ai <fangzhou.ai@amd.com>
Signed-off-by: fai <fangzhouai@gmail.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Cursor Grok 4.6 <cursoragent@cursor.com>
This commit is contained in:
Fangzhou Ai
2026-08-16 12:16:46 -07:00
committed by GitHub
co-authored by Nick Hill Cursor Grok 4.6
parent 1f0e0bf612
commit ef43e3101b
5 changed files with 1550 additions and 94 deletions
+48 -3
View File
@@ -19,6 +19,7 @@ from vllm.models.deepseek_v4.sparse_mla import (
DeepseekV4SparseMLAMetadataBuilder,
)
from vllm.platforms import current_platform
from vllm.platforms.rocm import _ON_GFX950
from vllm.triton_utils import tl, triton
from vllm.v1.attention.backend import (
CommonAttentionMetadata,
@@ -36,6 +37,19 @@ from vllm.v1.attention.ops.rocm_aiter_mla_sparse import (
from vllm.v1.worker.workspace import current_workspace_manager
def _trust_dsv4_extra_cache_nan_free(
kv_cache_dtype: str,
has_kv_transfer: bool,
has_extra_cache: bool,
) -> bool:
return (
_ON_GFX950
and kv_cache_dtype == "fp8_ds_mla"
and not has_kv_transfer
and has_extra_cache
)
def _build_indptr_from_lengths(lengths: torch.Tensor) -> torch.Tensor:
lengths = lengths.to(dtype=torch.int32).contiguous()
indptr = torch.zeros(lengths.shape[0] + 1, dtype=torch.int32, device=lengths.device)
@@ -294,9 +308,13 @@ def _copy_ragged_to_graph_buffers(
max_entries = max(num_rows * max_entries_per_row, 1)
ragged_out = ragged_indices_buffer[:max_entries]
nnz = ragged_indices.numel()
if nnz > 0:
ragged_out[:nnz].copy_(ragged_indices, non_blocking=True)
source_entries = ragged_indices.numel()
if source_entries > 0:
ragged_out[:source_entries].copy_(ragged_indices, non_blocking=True)
if _ON_GFX950:
# Preserve the graph-stable base pointer while exposing source capacity
# to the sync-free split selector; indptr still carries the true NNZ.
ragged_out = ragged_out[: max(source_entries, 1)]
return ragged_out, indptr_out
@@ -306,6 +324,7 @@ class DeepseekV4ROCMAiterMLASparseMetadata(DeepseekV4FlashMLAMetadata):
c128a_decode_topk_ragged_indices: torch.Tensor | None = None
c128a_decode_topk_ragged_indptr: torch.Tensor | None = None
for_cudagraph_capture: bool = False
@dataclass
@@ -370,6 +389,16 @@ class DeepseekV4ROCMAiterMLASparseMetadataBuilder(DeepseekV4SparseMLAMetadataBui
c128a_decode_topk_ragged_indptr=ragged_indptr,
)
def build_for_cudagraph_capture(
self, common_attn_metadata: CommonAttentionMetadata
) -> DeepseekV4ROCMAiterMLASparseMetadata:
metadata = cast(
DeepseekV4ROCMAiterMLASparseMetadata,
super().build_for_cudagraph_capture(common_attn_metadata),
)
metadata.for_cudagraph_capture = _ON_GFX950
return metadata
class DeepseekV4ROCMAiterSparseSWAMetadataBuilder(DeepseekSparseSWAMetadataBuilder):
# Keep fused multi-step decode disabled until update_draft_decode_metadata()
@@ -451,7 +480,9 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
backend_cls = DeepseekV4ROCMAiterMLASparseBackend
def __init__(self, *args, **kwargs):
vllm_config = args[0] if args else kwargs["vllm_config"]
super().__init__(*args, **kwargs)
self._has_kv_transfer = vllm_config.kv_transfer_config is not None
# Block scale for the preshuffled weight; None = not preshuffled.
self._wqa_wkv_scale: torch.Tensor | None = None
self._wo_b_scale: torch.Tensor | None = None
@@ -605,6 +636,13 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
attn_metadata=rocm_metadata,
swa_only=swa_only,
output=output[:num_decode_tokens],
adaptive_splits=(
_ON_GFX950
and not swa_only
and self.compress_ratio == 128
and rocm_metadata is not None
and rocm_metadata.for_cudagraph_capture
),
)
def _forward_decode(
@@ -615,6 +653,7 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
attn_metadata: DeepseekV4ROCMAiterMLASparseMetadata | None,
swa_only: bool,
output: torch.Tensor,
adaptive_splits: bool,
) -> None:
num_decodes = swa_metadata.num_decodes
num_decode_tokens = swa_metadata.num_decode_tokens
@@ -666,6 +705,12 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
nope_head_dim=self.nope_head_dim,
rope_head_dim=self.rope_head_dim,
output=output,
adaptive_splits=adaptive_splits,
extra_cache_nan_free=_trust_dsv4_extra_cache_nan_free(
self.kv_cache_dtype,
self._has_kv_transfer,
not swa_only and kv_cache is not None,
),
)
def _forward_prefill(
@@ -24,8 +24,14 @@ from typing import Any
import torch
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
if current_platform.is_rocm():
from vllm.platforms.rocm import _ON_GFX950
else:
_ON_GFX950 = False
from .fused_indexer_q import _fp32x2_to_fp4x2
@@ -61,12 +67,15 @@ def compress_norm_rope_store_triton(
if head_dim == 512:
kernel = _fused_kv_compress_norm_rope_insert_sparse_attn
num_warps = 4
kernel_kwargs = {"SANITIZE_CACHE_NANS": _ON_GFX950}
elif use_fp4_cache:
kernel = _fused_kv_compress_norm_rope_insert_indexer_mxfp4_attn
num_warps = 1
kernel_kwargs = {}
else:
kernel = _fused_kv_compress_norm_rope_insert_indexer_attn
num_warps = 1
kernel_kwargs = {}
kernel[(num_actual,)](
# state cache
@@ -103,6 +112,7 @@ def compress_norm_rope_store_triton(
SCALE_DIM=scale_dim,
KV_BLOCK_STRIDE=kv_cache.stride(0),
num_warps=num_warps,
**kernel_kwargs,
**pdl_kwargs,
)
@@ -145,6 +155,7 @@ def _fused_kv_compress_norm_rope_insert_sparse_attn(
TOKEN_STRIDE: tl.constexpr, # 576 for DeepseekV4
SCALE_DIM: tl.constexpr, # 8 for DeepseekV4 (7 real + 1 pad)
KV_BLOCK_STRIDE: tl.constexpr,
SANITIZE_CACHE_NANS: tl.constexpr,
):
"""Fused compress → RMSNorm → FP8 quant (nope) → RoPE → bf16 store (rope).
@@ -261,7 +272,8 @@ def _fused_kv_compress_norm_rope_insert_sparse_attn(
scale_idx = tl.arange(0, N_QUANT_BLOCKS)
encoded = exponents + 127.0
encoded = tl.maximum(tl.minimum(encoded, 255.0), 0.0)
max_encoded: tl.constexpr = 254.0 if SANITIZE_CACHE_NANS else 255.0
encoded = tl.maximum(tl.minimum(encoded, max_encoded), 0.0)
tl.store(
scale_ptr + scale_idx,
encoded.to(tl.uint8),
@@ -289,6 +301,8 @@ def _fused_kv_compress_norm_rope_insert_sparse_attn(
new_even = even * cos_v - odd * sin_v
new_odd = odd * cos_v + even * sin_v
result = tl.interleave(new_even, new_odd) # [TRITON_BLOCK_SIZE] fp32
if SANITIZE_CACHE_NANS:
result = tl.where(result == result, result, 0.0)
# Store rotated rope portion as bf16 into the cache's bf16 area.
bf16_ptr = (fp8_ptr + NOPE_HEAD_DIM).to(tl.pointer_type(tl.bfloat16))
@@ -417,6 +431,7 @@ def _finalize_norm_rope_quant_store_sparse_attn(
TOKEN_STRIDE: tl.constexpr,
SCALE_DIM: tl.constexpr,
KV_BLOCK_STRIDE: tl.constexpr,
SANITIZE_CACHE_NANS: tl.constexpr,
):
"""Stage 2: read compressed_kv[512] from scratch buffer, then
RMSNorm + FP8 quant (nope) + RoPE + bf16 store
@@ -474,7 +489,8 @@ def _finalize_norm_rope_quant_store_sparse_attn(
tl.store(fp8_ptr + block, x_uint8, mask=block < NOPE_HEAD_DIM)
scale_idx = tl.arange(0, N_QUANT_BLOCKS)
encoded = tl.maximum(tl.minimum(exponents + 127.0, 255.0), 0.0)
max_encoded: tl.constexpr = 254.0 if SANITIZE_CACHE_NANS else 255.0
encoded = tl.maximum(tl.minimum(exponents + 127.0, max_encoded), 0.0)
tl.store(
scale_ptr + scale_idx, encoded.to(tl.uint8), mask=scale_idx < N_NOPE_BLOCKS
)
@@ -494,6 +510,8 @@ def _finalize_norm_rope_quant_store_sparse_attn(
new_even = even * cos_v - odd * sin_v
new_odd = odd * cos_v + even * sin_v
result = tl.interleave(new_even, new_odd)
if SANITIZE_CACHE_NANS:
result = tl.where(result == result, result, 0.0)
bf16_ptr = (fp8_ptr + NOPE_HEAD_DIM).to(tl.pointer_type(tl.bfloat16))
rope_local = block - NOPE_HEAD_DIM
is_rope = (block >= NOPE_HEAD_DIM) & mask
@@ -564,6 +582,7 @@ def _launch_two_stage_sparse_attn_compressor(
TOKEN_STRIDE=token_stride,
SCALE_DIM=scale_dim,
KV_BLOCK_STRIDE=kv_cache.stride(0),
SANITIZE_CACHE_NANS=_ON_GFX950,
)